A band Mura detection method based on a new generative adversarial network

Defined as a visual inhomogeneity, Mura can cause seriously unpleasant feelings and that's why it needs to be inspected. Band Mura, which has a large area, is particularly difficult to be detected because of its irregular shape and size as well as its low contrast. So we propose a UADD-GAN model to detect band Mura in this work. Consisting of a proposed UADD generator and a discriminator, the model is trained with some normal samples, after which the generator is able to simulate the distribution of normal samples. During training, the generator takes normal images as inputs and output their reconstructions, while the discriminator receives images and determines whether they're true or fake, defiantly helping the generator to perform reconstructions better. The symmetric structure and operation of skip-adding make it easy for the UADD generator to reconstruct the normal samples well. On the contrary, the generator performs worse in the reconstruction of Mura samples so that we can distinguish them from the normal ones. In addition, we use multiple feature layers of the discriminator for supervision instead of using only the classification layer, helping the generator to reconstruct normal samples better. We've conducted experiments of MURA data sets with different proportion and our research indicates that the proposed model surpass other state of the art methods.